feat: ✨ Handle tool response checks with nemo guardrails check

[evaline-ju]

Summary

• Update ext-proc handling for checking for response bodies, which may be chunked. Here the strategy is to buffer chunks until they are all available to run checks on. Chunks may be empty and/or contain just end-of-stream (EOS) information
  • Add some initial server function tests, mainly around response body buffering and calling the post tool invocation hook
• Implement tool_post_invoke hook for Nemo guardrails check server, essentially calling the /guardrail/checks endpoint with role:tool and content ref. https://github.com/trustyai-explainability/trustyai-guardrails-operator/blob/main/docs/nemo_guardrails_quickstart.md
  • We focus on text for now, ignoring other potential content types, and multiple text arrays (if available) become concatenated
  • Add unit tests for this along with an addition for the existing tool_pre_invoke function
• Add line length on format/lint (can be adjusted to later to separate pyproject.toml etc.) and invoked this on files
• Makes errors more consistent for feature: Propagate error information from nemo check #19

Example call:

Related issue(s)

Fixes #18

[evaline-ju]

Planning to rebase on #34 so that updated tool response check logic does not have to go into two plugins

[julianstephen]

Lgtm. We can work on the tool response format within the other issue.